Synthetic sample · zero customer data · zero model calls

← proposed service

A run failed.
Do not blindly retry it.

This fictional packet contains a three-node n8n-style workflow and four synthetic log lines. The workflow loops AI Agent → Check Result → AI Agent. The logs show four attempts whose context grows from 9,000 to 35,000 tokens, and the same event ID appears twice.

Evidence inventory

Method: static bounded inspection only. No code execution, network access, or model call.

Finding 1 · model/retry cycle

Observed: the supplied graph contains a cycle joining the AI Agent and result-check node; the AI Agent is retry-marked.

Interpretation: this can amplify model calls when termination is weak. A cycle alone is not declared a defect.

Smallest check: make the loop exit/budget condition explicit and preserve retry state outside the growing prompt.

Acceptance: the agreed fixture reaches a terminal branch within the agreed maximum attempts and model-call cap.

Finding 2 · duplicate event identity

Observed: event ID msg-42 appears more than once in the supplied execution evidence.

Interpretation: whether it caused a duplicate external mutation is still unknown.

Smallest check: inspect durable idempotency/action receipts and the external target state before replay.

Acceptance: replaying a previously accepted event produces no second mutation or outward action.

Finding 3 · growing retry context

Observed: context telemetry grows 9000 → 15000 → 24000 → 35000.

Interpretation: if these entries are successive retries, transcript replay may be inflating model cost.

Smallest check: persist compact failure state and resend only minimum retry context.

Acceptance: a repeated failing fixture does not monotonically grow prompt context beyond the agreed bound.

Retry-safety receipt

HOLD_PRODUCTION_RETRY

Safe to blindly retry production: false.

Before another production run, verify durable action/idempotency receipts, the external target state for any possibly completed side effect, a bounded retry/model-call cap, and terminal-state evidence on a non-mutating fixture.

Allowed next step: repair or reproduce on synthetic/non-mutating evidence first.

This sample proves only the report shape on synthetic evidence. It is not a client result, production certification, or claim that every incident can be resolved statically.